OUTLINE - Stata · Ben A. Dwamena, MD Department of Radiology (Nuclear Medicine) Thursday October 25, 2007 FIRST WEST COAST STATA USERS' GROUP MEETING OUTLINE Evaluation of Diagnostic

1

Thursday October 25, 2007 FIRST WEST COAST STATA USERS' GROUP MEETING

METAANALYTIC INTEGRATION OF DIAGNOSTIC ACCURACY

STUDIES IN STATA

Ben A. Dwamena, MDDepartment of Radiology (Nuclear Medicine)


OUTLINE

Evaluation of Diagnostic TestsRationale For Meta-analysisExisting Methodology and LimitationsEmerging alternativeImplementation in Stata using midas

2


CLINICAL EVALUATION OF A

DIAGNOSTIC TEST

Consider a population to be made up of two groups: those with a disease and those without it. A test aims to identify people as belonging to one of these two groups.

It is usually assumed that a gold standard is available which can perfectly distinguish groups but cannot be used in routine practice due to problems such as invasiveness and/or cost.

However alternative, more practical, tests are available which are imperfect.


CLINICAL EVALUATION OF A DIAGNOSTIC TEST

Many test outcomes are measured on an explicit continuous scale (e.g. level of a chemical in the blood).

Changing the threshold test level which defines a positive and a negative test will also change the performance characteristics of a test.

3


THESHOLD EFFECT AND ROC


VARIATION IN THRESHOLD VS ACCURACY

4


2 X 2 TABLE FOR SINGLE STUDY


WHY META-ANALYSIS?

5


WHY META-ANALYSIS?


STATISTICAL ISSUES

Summary Estimates

Threshold Variability

Unobserved Heterogeneity

Covariate effects

Publication Bias

Clinical Interpretation

6


TRADITIONAL MODELSummary ROC Analysis and Regression


EXAMPLE DATAPeak Flow Velocity To Detect Renal Artery Stenosis

7


DISPLAY OF DATA IN ROC SPACE


SROC REGRESSION

8


SROC REGRESSION


LIMITATIONS OF SROC MODEL

9


BIVARIATE MIXED-EFFECTS BINOMIAL REGRESSION MODEL



10





11





12


BINOMIAL REGRESSION MODELXTMELOGIT CODE


BINOMIAL REGRESSION MODEL

GLAMM CODE

13


EXAMPLE DATASET:Axillary PET Scan in Primary Breast Cancer


SYNTAX FOR midasmidas varlist [if exp] [in range] [,

id(varname) year(varname) modeling_optionsquality_assessment_optionsreporting_optionsexploratory_graphics_optionspublication_bias_optionsforest_plot_optionsheterogeneity_optionsroc_optionsprobability_revision_optionsgeneral_graphing_options *]

14


MODELING Options

estimator(g|x)provides a choice between estimation with xtmelogit in release 10 versus gllamm in version 9 or earlier.

nip()specifies the number of integration points used for maximum likelihood estimation based on adaptive gaussian quadrature.

Default is set at 15 for midas (default in xtmelogit is 7).

Higher values improve accuracy at the expense of execution times.

model will be estimated by Laplacian approximation using nip(1)


QUALITY ASSESSMENT Options

qualitab

creates, using optional varlist of study quality items (presence=1, other=0) a table showing frequency of methodologic quality items.

qualibar

creates, combined with optional varlist of study quality items (presence=1, other=0) calculates study-specific quality scores and plots a bargraph of methodologic quality.

Qlab

may be combined with qualitab or qualibar to use variable labels for table and bargraph of methodologic items.

15


QUALITY ASSESSMENTmidas tp fp fn tn prodesign ssize30 fulverif testdescr /// refdescr subjdescr report brdspect blinded, qualib

0 20 40 60 80 100percent

blinded

brdspect

fulverif

prodesign

refdescr

report

ssize30

subjdescr

testdescr

BAR GRAPH OF QUALITY ASSESSMENT

Yes No


EXPLORATORY GRAPHICS Options

qqplot(dss|dlor|dlr)plots a normal quantile plot to (a) check the normality assumption (b) investigate whether all studies come from a single population (c) search for publication bias

cumproduces a cumulative meta-analysis plot using year of publication as measure for temporal evolution of evidence.

inf investigates the influence of each individual study on the overall meta-analysis summary estimate.

16


EXPLORATORY GRAPHICS Options

bivboximplements a two-dimensional analogue of the boxplot for univariate data It is used to assess distributional properties of sensitivity versus specificity and for indentifying possible outliers.

chiplot creates a chiplot for judging whether or not the paired performance indices are independent by augmenting the scatterplot with an auxiliary display.


CHI-PLOTmidas tp fp fn tn, id(author) year(year) ms(0.75) chip

17


BIVARIATE BOXPLOTmidas tp fp fn tn, id(author) year(year) ms(0.75) bivbox


REPORTING Options

table(dss|dlr|dlor)Creates a table of study specific performance estimates

Reports measure-specific summary estimates

Provides results of homogeneity (chi_squared) and inconsistency(I_squared) tests.

dss, dlr or dlor represent the paired performance measures:sensitivity/specificity, positive/negative likelihood ratios and diagnostic

score/odds ratios.

18


SUMMARY RESULTSmidas tp fp fn tn, id(author) year(year) es(x) ms(0.75) res(all)


REPORTING Options

results(all)

provides summary statistics, group-specific between-study variances, likelihood ratio test statistics and other global homogeneity tests.

results(het)

provides group-specific between-study variances, likelihood ratio test statistics and other global homogeneity tests.

results(sum)

provides summary statistics for all performance indices

19


SUMMARY RESULTSmidas tp fp fn tn, id(author) year(year) ms(0.75) es(x) res(sum)


PUBLICATION BIAS Optionspubbias

performs linear regression of log odds ratios on inverse root of effective sample sizes as a test for funnel plot asymmetry in diagnostic metanalyses. A non-zero slope coefficient is suggestive of significant small study bias (pvalue < 0.10).

maxbiasperforms Copas' worst-case sensitivity analysis for publication bias. calculates the upper limit of no of missing studies that will overturn statistical significance estimates the minimun likely publication probability (Copas and Jackson, 2004).

funnel plots a funnel plot, a two-dimensional graph with sample size on one axis and effect-size estimate on the other axis. The funnel plot capitalizes on the well-known statistical principle that sampling error decreases as sample size increases.

20


FUNNEL PLOTmidas tp fp fn tn, ms(0.75) fun


FUNNEL PLOT ASYMMETRY TESTmidas tp fp fn tn, pubb

21


FOREST PLOT Options

forest(dss|dlr|dlor)

Creates summary graphs with study-specific(box) and overall(diamond) point estimates and confidence intervals for each performance index pair

Confidence intervals lines are allowed to extend between 0 and 1000 beyond which they are truncated and marked by a leading arrow.

fordata

adds study-specific performance estimates and 95% CIs to right y-axis.


FOREST PLOTmidas tp fp fn tn, id(author) year(year) es(x) ms(0.75) ford for(dlr)

DLR POSITIVE (95% CI)

27.28[9.81 - 75.84]

1.80 [0.81 - 3.99]7.56 [1.68 - 34.04]2.39 [1.03 - 5.56]7.60 [1.69 - 34.19]5.89 [2.62 - 13.23]1.91 [0.90 - 4.05]15.55 [1.04 - 232.22]83.52 [5.26 - 1000.00]26.42 [1.69 - 414.06]9.78 [2.09 - 45.72]8.33 [1.84 - 37.69]14.25 [3.72 - 54.57]5.02 [2.52 - 10.00]7.18 [2.24 - 23.03]31.57 [2.04 - 488.19]3.34 [1.64 - 6.80]3.44 [1.90 - 6.24]18.05 [6.90 - 47.22]9.35 [4.53 - 19.32]4.00 [1.48 - 10.84]4.00 [1.48 - 10.84]

STUDY(YEAR)

Q = 67.63, df = 19.00, p = 0.00I2 = 59.30 [59.30 - 84.51]

COMBINED

Tse 1992Adler1 1993

Hoh 1993Crowe 1994

Avril 1996Bassa 1996

Scheidhauer 1996Utech 1996Adler2 1997

Palmedo 1997Noh 1998

Smith 1998Rostom 1999Yutani1 1999Hubner 2000

Ohta 2000Yutani2 2000

Greco 2001Schirrmeister 2001

Yang 2001

0.8 1000.0DLR POSITIVE

22


FOREST PLOTmidas tp fp fn tn, id(author) year(year) es(x) ms(0.45) for(dlr)

STUDY(YEAR)

Q = 40.28, df = 19.00, p = 0.00I2 = 52.84 [28.82 - 76.85]

COMBINEDTse 1992

Adler1 1993Hoh 1993

Crowe 1994Avril 1996

Bassa 1996Scheidhauer 1996

Utech 1996Adler2 1997

Palmedo 1997Noh 1998

Smith 1998Rostom 1999Yutani1 1999Hubner 2000

Ohta 2000Yutani2 2000

Greco 2001Schirrmeister 2001

Yang 2001

0 3DLR NEGATIVE


HETEROGENEITY Optionsgalb(dss|dlr|dlor)

The standardized effect measure (e.g. for lnDOR, lnDOR/precision) is plotted (y-axis) against the inverse of the precision(x-axis). A regression line that goes through the origin is calculated, together with 95% boundaries (starting at +2 and -2 on the y-axis). Studies outside these 95% boundaries may be considered as outliers.

hetfor creates composite forest plot of all performance indices to provide a general view of variability.Confidence intervals lines are allowed to extend between 0 and 1000 beyond which they are truncated and marked by a leading arrow.

covarscombined with an optional varlist permits univariable metaregression analysis of one or multiple covariables.

23


UNIVARIABLE META-REGRESSION

midas tp fp fn tn prodesign brdspect blinded, covars


COMPOSITE FOREST PLOTmidas tp fp fn tn, id(author) year(year) ms(0.75) hetfor scheme(s2manual)

T se 1 9 9 2

A d le r1 1 9 9 3

H o h 1 9 93

C ro w e 1 9 9 4

A vr i l 1 9 9 6

B a s sa 1 9 9 6

ch ei d h a ue r 19 96

U t e ch 1 99 6

A d le r2 1 9 9 7

P a lm e d o 1 9 9 7

N o h 1 9 98

S m i th 1 9 9 8

R o s to m 1 9 9 9

Y u t an i1 1 9 9 9

H u b n er 2 00 0

O h t a 2 0 0 0

Y u t an i2 2 0 0 0

G re c o 2 0 01

ch ir r m ei s t er 2 00 1

Y a n g 20 0 1

D a n f o r t h 2 0 0 2

G u l l e r 2 0 02

K e le m e n 2 0 0 2

N a k am o t o 1 2 0 0 2

N a k am o t o 2 2 0 0 2

R i e be r 2 0 0 2

a n _ H o e ve n 2 0 0 2

B a r ra n g er 2 00 3

F eh r 20 04

I n ou e 2 0 0 4

L ov r i c s 2 0 0 4

W a h l 20 0 4

Z or n o za 2 0 0 4

W e i r 2 0 0 5

G i l -R e n d o 2 0 06

K u m a r 20 06

S t a dn ik 2 0 0 6

C h u n g 20 06

V e ro n es i 2 0 0 6

S t u d yI d

0 . 0 1 .0

S e n siti vi ty

0 . 3 1 .0

S p e cifi cit y

0 . 3 7 31 . 3

P o s it iv e L i ke lih o o d R a tio

0 . 0 2 .8

N e ga t iv e L ike l ih o od R a t io

0 . 2 2 39 8 2 . 8

D ia g n os t ic O dd s R a ti o

24


ROC Optionssroc1

plots observed datapoints, summary operating sensitivity and specificity in SROC space.

sroc2

adds confidence and prediction contours.

rocplane

plots observed data in receiver operating characteristic space (ROC Plane) for visual assessment of threshold effect.


SUMMARY ROCmidas tp fp fn tn, es(x) ms(0.75) plot sroc1

25


SUMMARY ROCmidas tp fp fn tn, es(x) ms(0.75) plot sroc2


ROC PLANEmidas tp fp fn tn, id(author) year(year) es(x) rocp

26


PROBABILITY REVISION Options

pddam(p|r)produces a line graph of post-test probalities versus prior probabilities between 0 and 1 using summary likelihood ratios

lrmatrixcreates a scatter plot of positive and negative likelihood ratios with combined summary point. Plot is divided into quadrants based on strength-of-evidence thresholds to determine informativeness of measured test.


CONDITIONAL PROBABILITY PLOTmidas tp fp fn tn, es(x) pddam(p)

27


UNCONDITIONAL PREDICTIVE VALUESmidas tp fp fn tn, ms(0.75) es(x) pddam(r)


LIKELIHOOD RATIO SCATTERGRAMmidas tp fp fn tn, es(x) lrmat scheme(lean1)

28


PROBABILITY REVISION Options

fagan creates a plot showing the relationship between the prior probability, the likelihood ratio(combination of sensitivity andspecificity), and posterior test probability.

prior() combined with fagan allows user to specify a pretest probability overriding the default of using disease prevalence calculated from data when fagan is invoked alone.


FAGAN’S NOMOGRAM Imidas tp fp fn tn, es(x) fagan scheme(lean1)

29


FAGAN’S NOMOGRAM IImidas tp fp fn tn, es(x) fagan prior(.20) scheme(lean1)


STANDING ON SHOULDERS OF GIANTS

30


THANK YOU FOR YOUR ATTENTION

QUESTIONS?

SUGGESTIONS

COMMENTS


COMING SOON

New and Improved MIDAS package with routines for meta-analysis and meta-regression of roc curve area and continuous test result data.

Documents

OUTLINE - Stata · Ben A. Dwamena, MD Department of Radiology (Nuclear Medicine) Thursday October 25, 2007 FIRST WEST COAST STATA USERS' GROUP MEETING OUTLINE Evaluation of Diagnostic